Generating synthetic personal health data using conditional generative adversarial networks combining with...

Area of Science:

• Health Informatics
• Data Science
• Privacy-Preserving Technologies

Background:

• Personal health data is valuable but inaccessible due to privacy and legal constraints.
• Synthetic data offers a promising solution, but challenges remain in realism, privacy preservation, and handling imbalanced datasets.
• Existing methods struggle with minority class data simulation and capturing variable dependencies.

Purpose of the Study:

• To propose a novel differentially private conditional Generative Adversarial Network (DP-CGANS) model for generating realistic and privacy-preserving synthetic personal health data.
• To address the specific challenges of synthetic health data generation, including minority class representation and inter-variable dependency.
• To ensure a balance between data utility and patient privacy.

Main Methods:

• Developed a DP-CGANS model involving data transformation, sampling, conditioning, and network training.
• Distinguished and transformed categorical and continuous variables into latent space separately.
• Incorporated a conditional vector to represent minority classes and injected noise into gradients for differential privacy.

Main Results:

• The DP-CGANS model demonstrated superior performance in capturing variable dependencies compared to state-of-the-art models.
• Evaluations on socio-economic and real-world health datasets showed strong statistical similarity, machine learning utility, and privacy preservation.
• The model effectively balances data utility and privacy, even with imbalanced classes, abnormal distributions, and data sparsity.

Conclusions:

• DP-CGANS is an effective method for generating high-fidelity, privacy-preserving synthetic personal health data.
• The model successfully overcomes key challenges in synthetic health data generation, particularly for imbalanced and complex datasets.
• This approach enhances the accessibility of valuable health data for research while upholding stringent privacy standards.